Developing an AI model that links brain activity and disease

A research team from Baylor College of Medicine and Yale University incorporated generative artificial intelligence (AI) to create a basic model of brain activity. The Brain Language Model (BrainLM) was developed to model the brain in computers and determine how brain activity is related to human behavior and brain diseases. This research was presented as a conference paper at ICLR 2024, an influential venue in the field of deep learning AI.

“We've known for a long time that brain activity is related to human behavior and many diseases, including seizures and Parkinson's disease,” said Chadi Abdallah, associate professor in the Menninger Department of Psychiatry and Behavioral Sciences at Baylor University. says the doctor. Co-corresponding author of the paper. “While functional brain imaging and functional MRI allow us to observe brain activity throughout the brain, traditional data analysis tools cannot fully capture the dynamics of these activities in time and space. More recently, people have been able to use machine learning to capture the complexity of the brain and how it relates to certain diseases, but to It turned out to be a very expensive process, requiring thousands of patients with the disease to be enrolled and thoroughly tested.”

The power of new generative AI tools is that they can be used to create fundamental models that are independent of specific tasks or specific patient populations. Generative AI acts as a detective, uncovering hidden patterns within datasets. By analyzing data points and the relationships between them, these models can learn about the underlying dynamics: how and why things change or evolve. These basic models are fine-tuned to understand different topics. Researchers used generative AI to capture how brain activity functions regardless of a specific disorder or disease. It can be applied to any population without requiring knowledge of the subject's behavior, illness, medical history, or age. Brain activity is all that is needed to teach computers and AI models how brain activity evolves over space and time.

The team acquired 80,000 scans from 40,000 subjects and trained a model to understand how these brain activities are interconnected over time, creating the BrainLM Foundational Model of Brain Activity. has been established. Researchers can now use her BrainLM to fine-tune specific tasks or ask other research questions.

“For example, if you want to conduct a clinical trial to develop a drug to treat depression, you need to enroll large numbers of patients and treat them over long periods of time, which can cost hundreds of millions of dollars. By leveraging the power of BrainLM, we could potentially cut this cost in half by enrolling only half of the subjects and selecting those who would most benefit from the treatment. BrainLM can apply the knowledge learned from 80,000 scans to a specific subject.'' Abdallah said.

The first step, preprocessing, summarizes the signal and removes noise unrelated to brain activity. The researchers fed the summaries into a machine learning model and masked part of each person's data. Once the model learned the dynamics, it was tested on an excluded test group. They also tested this on different samples to understand how well the model generalizes to data acquired with different scanners and different populations, such as older and younger people. They found that BrainLM performed well on a variety of samples. We also found that BrainLM could more accurately predict the severity of depression, anxiety, and PTSD than other machine learning tools that don't use generative AI.

“We found that BrainLM worked very well, predicting brain activity in new samples that were hidden during training, and also performing well on data from new scanners and new populations.” Abdallah said. “These impressive results were achieved with scans from 40,000 subjects. We are currently working on significantly increasing our training dataset. The more powerful the models we can build, the more , more can be done to support patient care, such as the development of new treatments for mental illness, or neurosurgical guidance for seizures and DBS. ”

The researchers plan to apply the model to studies that predict future brain-related diseases.

Josue Ortega Caro, Antonio Enrique de Oliveira Fonseca, Syed A. Rizvi, Matteo Rosati, Christopher Averill, James L. Cross, Prateek Mittal, Emanuele Zappala, Rahul Madhav Dodapkar, David van Dijk also contributed to this work. The authors are affiliated with Baylor College of Medicine, Yale University, University of Southern California, and Idaho State University.

The study was funded by the Wu Tsai Institute at Yale University and the Beth K. Stuart Yudofsky Professor of Military Post-Traumatic Stress Syndrome Neuropsychiatry at Baylor University.